June 2026 summary for mirage-project/mirage: Delivered EAGLE3 speculative decoding for Qwen3 with parallel token generation for multi-slot decoding (K>1), including end-to-end correctness tests and a refactor into reusable components to improve maintainability. Fixed critical K>=2 path issues in the megakernel, resolved slot handling and MMA accumulator bugs, and aligned default parameters with template-driven design (MAX_TOKENS). Expanded test coverage with correctness harnesses and bf16-tie robustness checks, achieving exact matches for K=1 against the greedy reference and stabilized performance for K=2..5. Overall impact includes higher throughput for multi-user inference, stronger correctness guarantees, and a cleaner, scalable codebase ready for future model variants.